1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/kmemleak.h>
15 #include <linux/init.h>
16 #include <linux/slab.h>
17 #include <linux/workqueue.h>
18 #include <linux/smp.h>
19 #include <linux/llist.h>
20 #include <linux/list_sort.h>
21 #include <linux/cpu.h>
22 #include <linux/cache.h>
23 #include <linux/sched/sysctl.h>
24 #include <linux/sched/topology.h>
25 #include <linux/sched/signal.h>
26 #include <linux/delay.h>
27 #include <linux/crash_dump.h>
28 #include <linux/prefetch.h>
29 #include <linux/blk-crypto.h>
31 #include <trace/events/block.h>
33 #include <linux/blk-mq.h>
34 #include <linux/t10-pi.h>
37 #include "blk-mq-debugfs.h"
38 #include "blk-mq-tag.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
44 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
46 static void blk_mq_poll_stats_start(struct request_queue *q);
47 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
49 static int blk_mq_poll_stats_bkt(const struct request *rq)
51 int ddir, sectors, bucket;
53 ddir = rq_data_dir(rq);
54 sectors = blk_rq_stats_sectors(rq);
56 bucket = ddir + 2 * ilog2(sectors);
60 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
61 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
67 * Check if any of the ctx, dispatch list or elevator
68 * have pending work in this hardware queue.
70 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
72 return !list_empty_careful(&hctx->dispatch) ||
73 sbitmap_any_bit_set(&hctx->ctx_map) ||
74 blk_mq_sched_has_work(hctx);
78 * Mark this ctx as having pending work in this hardware queue
80 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
81 struct blk_mq_ctx *ctx)
83 const int bit = ctx->index_hw[hctx->type];
85 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
86 sbitmap_set_bit(&hctx->ctx_map, bit);
89 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
90 struct blk_mq_ctx *ctx)
92 const int bit = ctx->index_hw[hctx->type];
94 sbitmap_clear_bit(&hctx->ctx_map, bit);
98 struct hd_struct *part;
99 unsigned int inflight[2];
102 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
103 struct request *rq, void *priv,
106 struct mq_inflight *mi = priv;
108 if (rq->part == mi->part)
109 mi->inflight[rq_data_dir(rq)]++;
114 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
116 struct mq_inflight mi = { .part = part };
118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
120 return mi.inflight[0] + mi.inflight[1];
123 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
124 unsigned int inflight[2])
126 struct mq_inflight mi = { .part = part };
128 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
129 inflight[0] = mi.inflight[0];
130 inflight[1] = mi.inflight[1];
133 void blk_freeze_queue_start(struct request_queue *q)
135 mutex_lock(&q->mq_freeze_lock);
136 if (++q->mq_freeze_depth == 1) {
137 percpu_ref_kill(&q->q_usage_counter);
138 mutex_unlock(&q->mq_freeze_lock);
140 blk_mq_run_hw_queues(q, false);
142 mutex_unlock(&q->mq_freeze_lock);
145 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
147 void blk_mq_freeze_queue_wait(struct request_queue *q)
149 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
151 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
153 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
154 unsigned long timeout)
156 return wait_event_timeout(q->mq_freeze_wq,
157 percpu_ref_is_zero(&q->q_usage_counter),
160 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
163 * Guarantee no request is in use, so we can change any data structure of
164 * the queue afterward.
166 void blk_freeze_queue(struct request_queue *q)
169 * In the !blk_mq case we are only calling this to kill the
170 * q_usage_counter, otherwise this increases the freeze depth
171 * and waits for it to return to zero. For this reason there is
172 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
173 * exported to drivers as the only user for unfreeze is blk_mq.
175 blk_freeze_queue_start(q);
176 blk_mq_freeze_queue_wait(q);
179 void blk_mq_freeze_queue(struct request_queue *q)
182 * ...just an alias to keep freeze and unfreeze actions balanced
183 * in the blk_mq_* namespace
187 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
189 void blk_mq_unfreeze_queue(struct request_queue *q)
191 mutex_lock(&q->mq_freeze_lock);
192 q->mq_freeze_depth--;
193 WARN_ON_ONCE(q->mq_freeze_depth < 0);
194 if (!q->mq_freeze_depth) {
195 percpu_ref_resurrect(&q->q_usage_counter);
196 wake_up_all(&q->mq_freeze_wq);
198 mutex_unlock(&q->mq_freeze_lock);
200 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
203 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
204 * mpt3sas driver such that this function can be removed.
206 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
208 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
210 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
213 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
216 * Note: this function does not prevent that the struct request end_io()
217 * callback function is invoked. Once this function is returned, we make
218 * sure no dispatch can happen until the queue is unquiesced via
219 * blk_mq_unquiesce_queue().
221 void blk_mq_quiesce_queue(struct request_queue *q)
223 struct blk_mq_hw_ctx *hctx;
227 blk_mq_quiesce_queue_nowait(q);
229 queue_for_each_hw_ctx(q, hctx, i) {
230 if (hctx->flags & BLK_MQ_F_BLOCKING)
231 synchronize_srcu(hctx->srcu);
238 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
241 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
244 * This function recovers queue into the state before quiescing
245 * which is done by blk_mq_quiesce_queue.
247 void blk_mq_unquiesce_queue(struct request_queue *q)
249 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
251 /* dispatch requests which are inserted during quiescing */
252 blk_mq_run_hw_queues(q, true);
254 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
256 void blk_mq_wake_waiters(struct request_queue *q)
258 struct blk_mq_hw_ctx *hctx;
261 queue_for_each_hw_ctx(q, hctx, i)
262 if (blk_mq_hw_queue_mapped(hctx))
263 blk_mq_tag_wakeup_all(hctx->tags, true);
267 * Only need start/end time stamping if we have iostat or
268 * blk stats enabled, or using an IO scheduler.
270 static inline bool blk_mq_need_time_stamp(struct request *rq)
272 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
275 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
276 unsigned int tag, u64 alloc_time_ns)
278 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
279 struct request *rq = tags->static_rqs[tag];
281 if (data->q->elevator) {
282 rq->tag = BLK_MQ_NO_TAG;
283 rq->internal_tag = tag;
286 rq->internal_tag = BLK_MQ_NO_TAG;
289 /* csd/requeue_work/fifo_time is initialized before use */
291 rq->mq_ctx = data->ctx;
292 rq->mq_hctx = data->hctx;
294 rq->cmd_flags = data->cmd_flags;
295 if (data->flags & BLK_MQ_REQ_PREEMPT)
296 rq->rq_flags |= RQF_PREEMPT;
297 if (blk_queue_io_stat(data->q))
298 rq->rq_flags |= RQF_IO_STAT;
299 INIT_LIST_HEAD(&rq->queuelist);
300 INIT_HLIST_NODE(&rq->hash);
301 RB_CLEAR_NODE(&rq->rb_node);
304 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
305 rq->alloc_time_ns = alloc_time_ns;
307 if (blk_mq_need_time_stamp(rq))
308 rq->start_time_ns = ktime_get_ns();
310 rq->start_time_ns = 0;
311 rq->io_start_time_ns = 0;
312 rq->stats_sectors = 0;
313 rq->nr_phys_segments = 0;
314 #if defined(CONFIG_BLK_DEV_INTEGRITY)
315 rq->nr_integrity_segments = 0;
317 blk_crypto_rq_set_defaults(rq);
318 /* tag was already set */
319 WRITE_ONCE(rq->deadline, 0);
324 rq->end_io_data = NULL;
326 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
327 refcount_set(&rq->ref, 1);
329 if (!op_is_flush(data->cmd_flags)) {
330 struct elevator_queue *e = data->q->elevator;
333 if (e && e->type->ops.prepare_request) {
334 if (e->type->icq_cache)
335 blk_mq_sched_assign_ioc(rq);
337 e->type->ops.prepare_request(rq);
338 rq->rq_flags |= RQF_ELVPRIV;
342 data->hctx->queued++;
346 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
348 struct request_queue *q = data->q;
349 struct elevator_queue *e = q->elevator;
350 u64 alloc_time_ns = 0;
353 /* alloc_time includes depth and tag waits */
354 if (blk_queue_rq_alloc_time(q))
355 alloc_time_ns = ktime_get_ns();
357 if (data->cmd_flags & REQ_NOWAIT)
358 data->flags |= BLK_MQ_REQ_NOWAIT;
362 * Flush requests are special and go directly to the
363 * dispatch list. Don't include reserved tags in the
364 * limiting, as it isn't useful.
366 if (!op_is_flush(data->cmd_flags) &&
367 e->type->ops.limit_depth &&
368 !(data->flags & BLK_MQ_REQ_RESERVED))
369 e->type->ops.limit_depth(data->cmd_flags, data);
373 data->ctx = blk_mq_get_ctx(q);
374 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
376 blk_mq_tag_busy(data->hctx);
379 * Waiting allocations only fail because of an inactive hctx. In that
380 * case just retry the hctx assignment and tag allocation as CPU hotplug
381 * should have migrated us to an online CPU by now.
383 tag = blk_mq_get_tag(data);
384 if (tag == BLK_MQ_NO_TAG) {
385 if (data->flags & BLK_MQ_REQ_NOWAIT)
389 * Give up the CPU and sleep for a random short time to ensure
390 * that thread using a realtime scheduling class are migrated
391 * off the CPU, and thus off the hctx that is going away.
396 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
399 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
400 blk_mq_req_flags_t flags)
402 struct blk_mq_alloc_data data = {
410 ret = blk_queue_enter(q, flags);
414 rq = __blk_mq_alloc_request(&data);
418 rq->__sector = (sector_t) -1;
419 rq->bio = rq->biotail = NULL;
423 return ERR_PTR(-EWOULDBLOCK);
425 EXPORT_SYMBOL(blk_mq_alloc_request);
427 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
428 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
430 struct blk_mq_alloc_data data = {
435 u64 alloc_time_ns = 0;
440 /* alloc_time includes depth and tag waits */
441 if (blk_queue_rq_alloc_time(q))
442 alloc_time_ns = ktime_get_ns();
445 * If the tag allocator sleeps we could get an allocation for a
446 * different hardware context. No need to complicate the low level
447 * allocator for this for the rare use case of a command tied to
450 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
451 return ERR_PTR(-EINVAL);
453 if (hctx_idx >= q->nr_hw_queues)
454 return ERR_PTR(-EIO);
456 ret = blk_queue_enter(q, flags);
461 * Check if the hardware context is actually mapped to anything.
462 * If not tell the caller that it should skip this queue.
465 data.hctx = q->queue_hw_ctx[hctx_idx];
466 if (!blk_mq_hw_queue_mapped(data.hctx))
468 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
469 data.ctx = __blk_mq_get_ctx(q, cpu);
472 blk_mq_tag_busy(data.hctx);
475 tag = blk_mq_get_tag(&data);
476 if (tag == BLK_MQ_NO_TAG)
478 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
484 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
486 static void __blk_mq_free_request(struct request *rq)
488 struct request_queue *q = rq->q;
489 struct blk_mq_ctx *ctx = rq->mq_ctx;
490 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
491 const int sched_tag = rq->internal_tag;
493 blk_crypto_free_request(rq);
494 blk_pm_mark_last_busy(rq);
496 if (rq->tag != BLK_MQ_NO_TAG)
497 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
498 if (sched_tag != BLK_MQ_NO_TAG)
499 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
500 blk_mq_sched_restart(hctx);
504 void blk_mq_free_request(struct request *rq)
506 struct request_queue *q = rq->q;
507 struct elevator_queue *e = q->elevator;
508 struct blk_mq_ctx *ctx = rq->mq_ctx;
509 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
511 if (rq->rq_flags & RQF_ELVPRIV) {
512 if (e && e->type->ops.finish_request)
513 e->type->ops.finish_request(rq);
515 put_io_context(rq->elv.icq->ioc);
520 ctx->rq_completed[rq_is_sync(rq)]++;
521 if (rq->rq_flags & RQF_MQ_INFLIGHT)
522 atomic_dec(&hctx->nr_active);
524 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
525 laptop_io_completion(q->backing_dev_info);
529 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
530 if (refcount_dec_and_test(&rq->ref))
531 __blk_mq_free_request(rq);
533 EXPORT_SYMBOL_GPL(blk_mq_free_request);
535 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
539 if (blk_mq_need_time_stamp(rq))
540 now = ktime_get_ns();
542 if (rq->rq_flags & RQF_STATS) {
543 blk_mq_poll_stats_start(rq->q);
544 blk_stat_add(rq, now);
547 blk_mq_sched_completed_request(rq, now);
549 blk_account_io_done(rq, now);
552 rq_qos_done(rq->q, rq);
553 rq->end_io(rq, error);
555 blk_mq_free_request(rq);
558 EXPORT_SYMBOL(__blk_mq_end_request);
560 void blk_mq_end_request(struct request *rq, blk_status_t error)
562 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
564 __blk_mq_end_request(rq, error);
566 EXPORT_SYMBOL(blk_mq_end_request);
569 * Softirq action handler - move entries to local list and loop over them
570 * while passing them to the queue registered handler.
572 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
574 struct list_head *cpu_list, local_list;
577 cpu_list = this_cpu_ptr(&blk_cpu_done);
578 list_replace_init(cpu_list, &local_list);
581 while (!list_empty(&local_list)) {
584 rq = list_entry(local_list.next, struct request, ipi_list);
585 list_del_init(&rq->ipi_list);
586 rq->q->mq_ops->complete(rq);
590 static void blk_mq_trigger_softirq(struct request *rq)
592 struct list_head *list;
595 local_irq_save(flags);
596 list = this_cpu_ptr(&blk_cpu_done);
597 list_add_tail(&rq->ipi_list, list);
600 * If the list only contains our just added request, signal a raise of
601 * the softirq. If there are already entries there, someone already
602 * raised the irq but it hasn't run yet.
604 if (list->next == &rq->ipi_list)
605 raise_softirq_irqoff(BLOCK_SOFTIRQ);
606 local_irq_restore(flags);
609 static int blk_softirq_cpu_dead(unsigned int cpu)
612 * If a CPU goes away, splice its entries to the current CPU
613 * and trigger a run of the softirq
616 list_splice_init(&per_cpu(blk_cpu_done, cpu),
617 this_cpu_ptr(&blk_cpu_done));
618 raise_softirq_irqoff(BLOCK_SOFTIRQ);
625 static void __blk_mq_complete_request_remote(void *data)
627 struct request *rq = data;
630 * For most of single queue controllers, there is only one irq vector
631 * for handling I/O completion, and the only irq's affinity is set
632 * to all possible CPUs. On most of ARCHs, this affinity means the irq
633 * is handled on one specific CPU.
635 * So complete I/O requests in softirq context in case of single queue
636 * devices to avoid degrading I/O performance due to irqsoff latency.
638 if (rq->q->nr_hw_queues == 1)
639 blk_mq_trigger_softirq(rq);
641 rq->q->mq_ops->complete(rq);
644 static inline bool blk_mq_complete_need_ipi(struct request *rq)
646 int cpu = raw_smp_processor_id();
648 if (!IS_ENABLED(CONFIG_SMP) ||
649 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
652 /* same CPU or cache domain? Complete locally */
653 if (cpu == rq->mq_ctx->cpu ||
654 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
655 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
658 /* don't try to IPI to an offline CPU */
659 return cpu_online(rq->mq_ctx->cpu);
662 bool blk_mq_complete_request_remote(struct request *rq)
664 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
667 * For a polled request, always complete locallly, it's pointless
668 * to redirect the completion.
670 if (rq->cmd_flags & REQ_HIPRI)
673 if (blk_mq_complete_need_ipi(rq)) {
674 rq->csd.func = __blk_mq_complete_request_remote;
677 smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
679 if (rq->q->nr_hw_queues > 1)
681 blk_mq_trigger_softirq(rq);
686 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
689 * blk_mq_complete_request - end I/O on a request
690 * @rq: the request being processed
693 * Complete a request by scheduling the ->complete_rq operation.
695 void blk_mq_complete_request(struct request *rq)
697 if (!blk_mq_complete_request_remote(rq))
698 rq->q->mq_ops->complete(rq);
700 EXPORT_SYMBOL(blk_mq_complete_request);
702 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
703 __releases(hctx->srcu)
705 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
708 srcu_read_unlock(hctx->srcu, srcu_idx);
711 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
712 __acquires(hctx->srcu)
714 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
715 /* shut up gcc false positive */
719 *srcu_idx = srcu_read_lock(hctx->srcu);
723 * blk_mq_start_request - Start processing a request
724 * @rq: Pointer to request to be started
726 * Function used by device drivers to notify the block layer that a request
727 * is going to be processed now, so blk layer can do proper initializations
728 * such as starting the timeout timer.
730 void blk_mq_start_request(struct request *rq)
732 struct request_queue *q = rq->q;
734 trace_block_rq_issue(q, rq);
736 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
737 rq->io_start_time_ns = ktime_get_ns();
738 rq->stats_sectors = blk_rq_sectors(rq);
739 rq->rq_flags |= RQF_STATS;
743 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
746 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
748 #ifdef CONFIG_BLK_DEV_INTEGRITY
749 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
750 q->integrity.profile->prepare_fn(rq);
753 EXPORT_SYMBOL(blk_mq_start_request);
755 static void __blk_mq_requeue_request(struct request *rq)
757 struct request_queue *q = rq->q;
759 blk_mq_put_driver_tag(rq);
761 trace_block_rq_requeue(q, rq);
762 rq_qos_requeue(q, rq);
764 if (blk_mq_request_started(rq)) {
765 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
766 rq->rq_flags &= ~RQF_TIMED_OUT;
770 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
772 __blk_mq_requeue_request(rq);
774 /* this request will be re-inserted to io scheduler queue */
775 blk_mq_sched_requeue_request(rq);
777 BUG_ON(!list_empty(&rq->queuelist));
778 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
780 EXPORT_SYMBOL(blk_mq_requeue_request);
782 static void blk_mq_requeue_work(struct work_struct *work)
784 struct request_queue *q =
785 container_of(work, struct request_queue, requeue_work.work);
787 struct request *rq, *next;
789 spin_lock_irq(&q->requeue_lock);
790 list_splice_init(&q->requeue_list, &rq_list);
791 spin_unlock_irq(&q->requeue_lock);
793 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
794 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
797 rq->rq_flags &= ~RQF_SOFTBARRIER;
798 list_del_init(&rq->queuelist);
800 * If RQF_DONTPREP, rq has contained some driver specific
801 * data, so insert it to hctx dispatch list to avoid any
804 if (rq->rq_flags & RQF_DONTPREP)
805 blk_mq_request_bypass_insert(rq, false, false);
807 blk_mq_sched_insert_request(rq, true, false, false);
810 while (!list_empty(&rq_list)) {
811 rq = list_entry(rq_list.next, struct request, queuelist);
812 list_del_init(&rq->queuelist);
813 blk_mq_sched_insert_request(rq, false, false, false);
816 blk_mq_run_hw_queues(q, false);
819 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
820 bool kick_requeue_list)
822 struct request_queue *q = rq->q;
826 * We abuse this flag that is otherwise used by the I/O scheduler to
827 * request head insertion from the workqueue.
829 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
831 spin_lock_irqsave(&q->requeue_lock, flags);
833 rq->rq_flags |= RQF_SOFTBARRIER;
834 list_add(&rq->queuelist, &q->requeue_list);
836 list_add_tail(&rq->queuelist, &q->requeue_list);
838 spin_unlock_irqrestore(&q->requeue_lock, flags);
840 if (kick_requeue_list)
841 blk_mq_kick_requeue_list(q);
844 void blk_mq_kick_requeue_list(struct request_queue *q)
846 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
848 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
850 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
853 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
854 msecs_to_jiffies(msecs));
856 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
858 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
860 if (tag < tags->nr_tags) {
861 prefetch(tags->rqs[tag]);
862 return tags->rqs[tag];
867 EXPORT_SYMBOL(blk_mq_tag_to_rq);
869 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
870 void *priv, bool reserved)
873 * If we find a request that isn't idle and the queue matches,
874 * we know the queue is busy. Return false to stop the iteration.
876 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
886 bool blk_mq_queue_inflight(struct request_queue *q)
890 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
893 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
895 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
897 req->rq_flags |= RQF_TIMED_OUT;
898 if (req->q->mq_ops->timeout) {
899 enum blk_eh_timer_return ret;
901 ret = req->q->mq_ops->timeout(req, reserved);
902 if (ret == BLK_EH_DONE)
904 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
910 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
912 unsigned long deadline;
914 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
916 if (rq->rq_flags & RQF_TIMED_OUT)
919 deadline = READ_ONCE(rq->deadline);
920 if (time_after_eq(jiffies, deadline))
925 else if (time_after(*next, deadline))
930 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
931 struct request *rq, void *priv, bool reserved)
933 unsigned long *next = priv;
936 * Just do a quick check if it is expired before locking the request in
937 * so we're not unnecessarilly synchronizing across CPUs.
939 if (!blk_mq_req_expired(rq, next))
943 * We have reason to believe the request may be expired. Take a
944 * reference on the request to lock this request lifetime into its
945 * currently allocated context to prevent it from being reallocated in
946 * the event the completion by-passes this timeout handler.
948 * If the reference was already released, then the driver beat the
949 * timeout handler to posting a natural completion.
951 if (!refcount_inc_not_zero(&rq->ref))
955 * The request is now locked and cannot be reallocated underneath the
956 * timeout handler's processing. Re-verify this exact request is truly
957 * expired; if it is not expired, then the request was completed and
958 * reallocated as a new request.
960 if (blk_mq_req_expired(rq, next))
961 blk_mq_rq_timed_out(rq, reserved);
963 if (is_flush_rq(rq, hctx))
965 else if (refcount_dec_and_test(&rq->ref))
966 __blk_mq_free_request(rq);
971 static void blk_mq_timeout_work(struct work_struct *work)
973 struct request_queue *q =
974 container_of(work, struct request_queue, timeout_work);
975 unsigned long next = 0;
976 struct blk_mq_hw_ctx *hctx;
979 /* A deadlock might occur if a request is stuck requiring a
980 * timeout at the same time a queue freeze is waiting
981 * completion, since the timeout code would not be able to
982 * acquire the queue reference here.
984 * That's why we don't use blk_queue_enter here; instead, we use
985 * percpu_ref_tryget directly, because we need to be able to
986 * obtain a reference even in the short window between the queue
987 * starting to freeze, by dropping the first reference in
988 * blk_freeze_queue_start, and the moment the last request is
989 * consumed, marked by the instant q_usage_counter reaches
992 if (!percpu_ref_tryget(&q->q_usage_counter))
995 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
998 mod_timer(&q->timeout, next);
1001 * Request timeouts are handled as a forward rolling timer. If
1002 * we end up here it means that no requests are pending and
1003 * also that no request has been pending for a while. Mark
1004 * each hctx as idle.
1006 queue_for_each_hw_ctx(q, hctx, i) {
1007 /* the hctx may be unmapped, so check it here */
1008 if (blk_mq_hw_queue_mapped(hctx))
1009 blk_mq_tag_idle(hctx);
1015 struct flush_busy_ctx_data {
1016 struct blk_mq_hw_ctx *hctx;
1017 struct list_head *list;
1020 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1022 struct flush_busy_ctx_data *flush_data = data;
1023 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1024 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1025 enum hctx_type type = hctx->type;
1027 spin_lock(&ctx->lock);
1028 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1029 sbitmap_clear_bit(sb, bitnr);
1030 spin_unlock(&ctx->lock);
1035 * Process software queues that have been marked busy, splicing them
1036 * to the for-dispatch
1038 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1040 struct flush_busy_ctx_data data = {
1045 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1047 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1049 struct dispatch_rq_data {
1050 struct blk_mq_hw_ctx *hctx;
1054 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1057 struct dispatch_rq_data *dispatch_data = data;
1058 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1059 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1060 enum hctx_type type = hctx->type;
1062 spin_lock(&ctx->lock);
1063 if (!list_empty(&ctx->rq_lists[type])) {
1064 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1065 list_del_init(&dispatch_data->rq->queuelist);
1066 if (list_empty(&ctx->rq_lists[type]))
1067 sbitmap_clear_bit(sb, bitnr);
1069 spin_unlock(&ctx->lock);
1071 return !dispatch_data->rq;
1074 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1075 struct blk_mq_ctx *start)
1077 unsigned off = start ? start->index_hw[hctx->type] : 0;
1078 struct dispatch_rq_data data = {
1083 __sbitmap_for_each_set(&hctx->ctx_map, off,
1084 dispatch_rq_from_ctx, &data);
1089 static inline unsigned int queued_to_index(unsigned int queued)
1094 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1097 static bool __blk_mq_get_driver_tag(struct request *rq)
1099 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1100 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1103 blk_mq_tag_busy(rq->mq_hctx);
1105 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1106 bt = &rq->mq_hctx->tags->breserved_tags;
1110 if (!hctx_may_queue(rq->mq_hctx, bt))
1112 tag = __sbitmap_queue_get(bt);
1113 if (tag == BLK_MQ_NO_TAG)
1116 rq->tag = tag + tag_offset;
1120 static bool blk_mq_get_driver_tag(struct request *rq)
1122 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1124 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1127 if ((hctx->flags & BLK_MQ_F_TAG_SHARED) &&
1128 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1129 rq->rq_flags |= RQF_MQ_INFLIGHT;
1130 atomic_inc(&hctx->nr_active);
1132 hctx->tags->rqs[rq->tag] = rq;
1136 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1137 int flags, void *key)
1139 struct blk_mq_hw_ctx *hctx;
1141 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1143 spin_lock(&hctx->dispatch_wait_lock);
1144 if (!list_empty(&wait->entry)) {
1145 struct sbitmap_queue *sbq;
1147 list_del_init(&wait->entry);
1148 sbq = &hctx->tags->bitmap_tags;
1149 atomic_dec(&sbq->ws_active);
1151 spin_unlock(&hctx->dispatch_wait_lock);
1153 blk_mq_run_hw_queue(hctx, true);
1158 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1159 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1160 * restart. For both cases, take care to check the condition again after
1161 * marking us as waiting.
1163 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1166 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1167 struct wait_queue_head *wq;
1168 wait_queue_entry_t *wait;
1171 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1172 blk_mq_sched_mark_restart_hctx(hctx);
1175 * It's possible that a tag was freed in the window between the
1176 * allocation failure and adding the hardware queue to the wait
1179 * Don't clear RESTART here, someone else could have set it.
1180 * At most this will cost an extra queue run.
1182 return blk_mq_get_driver_tag(rq);
1185 wait = &hctx->dispatch_wait;
1186 if (!list_empty_careful(&wait->entry))
1189 wq = &bt_wait_ptr(sbq, hctx)->wait;
1191 spin_lock_irq(&wq->lock);
1192 spin_lock(&hctx->dispatch_wait_lock);
1193 if (!list_empty(&wait->entry)) {
1194 spin_unlock(&hctx->dispatch_wait_lock);
1195 spin_unlock_irq(&wq->lock);
1199 atomic_inc(&sbq->ws_active);
1200 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1201 __add_wait_queue(wq, wait);
1204 * It's possible that a tag was freed in the window between the
1205 * allocation failure and adding the hardware queue to the wait
1208 ret = blk_mq_get_driver_tag(rq);
1210 spin_unlock(&hctx->dispatch_wait_lock);
1211 spin_unlock_irq(&wq->lock);
1216 * We got a tag, remove ourselves from the wait queue to ensure
1217 * someone else gets the wakeup.
1219 list_del_init(&wait->entry);
1220 atomic_dec(&sbq->ws_active);
1221 spin_unlock(&hctx->dispatch_wait_lock);
1222 spin_unlock_irq(&wq->lock);
1227 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1228 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1230 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1231 * - EWMA is one simple way to compute running average value
1232 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1233 * - take 4 as factor for avoiding to get too small(0) result, and this
1234 * factor doesn't matter because EWMA decreases exponentially
1236 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1240 if (hctx->queue->elevator)
1243 ewma = hctx->dispatch_busy;
1248 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1250 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1251 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1253 hctx->dispatch_busy = ewma;
1256 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1258 static void blk_mq_handle_dev_resource(struct request *rq,
1259 struct list_head *list)
1261 struct request *next =
1262 list_first_entry_or_null(list, struct request, queuelist);
1265 * If an I/O scheduler has been configured and we got a driver tag for
1266 * the next request already, free it.
1269 blk_mq_put_driver_tag(next);
1271 list_add(&rq->queuelist, list);
1272 __blk_mq_requeue_request(rq);
1275 static void blk_mq_handle_zone_resource(struct request *rq,
1276 struct list_head *zone_list)
1279 * If we end up here it is because we cannot dispatch a request to a
1280 * specific zone due to LLD level zone-write locking or other zone
1281 * related resource not being available. In this case, set the request
1282 * aside in zone_list for retrying it later.
1284 list_add(&rq->queuelist, zone_list);
1285 __blk_mq_requeue_request(rq);
1288 enum prep_dispatch {
1290 PREP_DISPATCH_NO_TAG,
1291 PREP_DISPATCH_NO_BUDGET,
1294 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1297 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1299 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1300 blk_mq_put_driver_tag(rq);
1301 return PREP_DISPATCH_NO_BUDGET;
1304 if (!blk_mq_get_driver_tag(rq)) {
1306 * The initial allocation attempt failed, so we need to
1307 * rerun the hardware queue when a tag is freed. The
1308 * waitqueue takes care of that. If the queue is run
1309 * before we add this entry back on the dispatch list,
1310 * we'll re-run it below.
1312 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1314 * All budgets not got from this function will be put
1315 * together during handling partial dispatch
1318 blk_mq_put_dispatch_budget(rq->q);
1319 return PREP_DISPATCH_NO_TAG;
1323 return PREP_DISPATCH_OK;
1326 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1327 static void blk_mq_release_budgets(struct request_queue *q,
1328 unsigned int nr_budgets)
1332 for (i = 0; i < nr_budgets; i++)
1333 blk_mq_put_dispatch_budget(q);
1337 * Returns true if we did some work AND can potentially do more.
1339 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1340 unsigned int nr_budgets)
1342 enum prep_dispatch prep;
1343 struct request_queue *q = hctx->queue;
1344 struct request *rq, *nxt;
1346 blk_status_t ret = BLK_STS_OK;
1347 LIST_HEAD(zone_list);
1349 if (list_empty(list))
1353 * Now process all the entries, sending them to the driver.
1355 errors = queued = 0;
1357 struct blk_mq_queue_data bd;
1359 rq = list_first_entry(list, struct request, queuelist);
1361 WARN_ON_ONCE(hctx != rq->mq_hctx);
1362 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1363 if (prep != PREP_DISPATCH_OK)
1366 list_del_init(&rq->queuelist);
1371 * Flag last if we have no more requests, or if we have more
1372 * but can't assign a driver tag to it.
1374 if (list_empty(list))
1377 nxt = list_first_entry(list, struct request, queuelist);
1378 bd.last = !blk_mq_get_driver_tag(nxt);
1382 * once the request is queued to lld, no need to cover the
1387 ret = q->mq_ops->queue_rq(hctx, &bd);
1392 case BLK_STS_RESOURCE:
1393 case BLK_STS_DEV_RESOURCE:
1394 blk_mq_handle_dev_resource(rq, list);
1396 case BLK_STS_ZONE_RESOURCE:
1398 * Move the request to zone_list and keep going through
1399 * the dispatch list to find more requests the drive can
1402 blk_mq_handle_zone_resource(rq, &zone_list);
1406 blk_mq_end_request(rq, BLK_STS_IOERR);
1408 } while (!list_empty(list));
1410 if (!list_empty(&zone_list))
1411 list_splice_tail_init(&zone_list, list);
1413 hctx->dispatched[queued_to_index(queued)]++;
1416 * Any items that need requeuing? Stuff them into hctx->dispatch,
1417 * that is where we will continue on next queue run.
1419 if (!list_empty(list)) {
1421 /* For non-shared tags, the RESTART check will suffice */
1422 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1423 (hctx->flags & BLK_MQ_F_TAG_SHARED);
1424 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1426 blk_mq_release_budgets(q, nr_budgets);
1429 * If we didn't flush the entire list, we could have told
1430 * the driver there was more coming, but that turned out to
1433 if (q->mq_ops->commit_rqs && queued)
1434 q->mq_ops->commit_rqs(hctx);
1436 spin_lock(&hctx->lock);
1437 list_splice_tail_init(list, &hctx->dispatch);
1438 spin_unlock(&hctx->lock);
1441 * Order adding requests to hctx->dispatch and checking
1442 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1443 * in blk_mq_sched_restart(). Avoid restart code path to
1444 * miss the new added requests to hctx->dispatch, meantime
1445 * SCHED_RESTART is observed here.
1450 * If SCHED_RESTART was set by the caller of this function and
1451 * it is no longer set that means that it was cleared by another
1452 * thread and hence that a queue rerun is needed.
1454 * If 'no_tag' is set, that means that we failed getting
1455 * a driver tag with an I/O scheduler attached. If our dispatch
1456 * waitqueue is no longer active, ensure that we run the queue
1457 * AFTER adding our entries back to the list.
1459 * If no I/O scheduler has been configured it is possible that
1460 * the hardware queue got stopped and restarted before requests
1461 * were pushed back onto the dispatch list. Rerun the queue to
1462 * avoid starvation. Notes:
1463 * - blk_mq_run_hw_queue() checks whether or not a queue has
1464 * been stopped before rerunning a queue.
1465 * - Some but not all block drivers stop a queue before
1466 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1469 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1470 * bit is set, run queue after a delay to avoid IO stalls
1471 * that could otherwise occur if the queue is idle. We'll do
1472 * similar if we couldn't get budget and SCHED_RESTART is set.
1474 needs_restart = blk_mq_sched_needs_restart(hctx);
1475 if (!needs_restart ||
1476 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1477 blk_mq_run_hw_queue(hctx, true);
1478 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1480 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1482 blk_mq_update_dispatch_busy(hctx, true);
1485 blk_mq_update_dispatch_busy(hctx, false);
1487 return (queued + errors) != 0;
1491 * __blk_mq_run_hw_queue - Run a hardware queue.
1492 * @hctx: Pointer to the hardware queue to run.
1494 * Send pending requests to the hardware.
1496 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1501 * We should be running this queue from one of the CPUs that
1504 * There are at least two related races now between setting
1505 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1506 * __blk_mq_run_hw_queue():
1508 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1509 * but later it becomes online, then this warning is harmless
1512 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1513 * but later it becomes offline, then the warning can't be
1514 * triggered, and we depend on blk-mq timeout handler to
1515 * handle dispatched requests to this hctx
1517 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1518 cpu_online(hctx->next_cpu)) {
1519 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1520 raw_smp_processor_id(),
1521 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1526 * We can't run the queue inline with ints disabled. Ensure that
1527 * we catch bad users of this early.
1529 WARN_ON_ONCE(in_interrupt());
1531 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1533 hctx_lock(hctx, &srcu_idx);
1534 blk_mq_sched_dispatch_requests(hctx);
1535 hctx_unlock(hctx, srcu_idx);
1538 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1540 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1542 if (cpu >= nr_cpu_ids)
1543 cpu = cpumask_first(hctx->cpumask);
1548 * It'd be great if the workqueue API had a way to pass
1549 * in a mask and had some smarts for more clever placement.
1550 * For now we just round-robin here, switching for every
1551 * BLK_MQ_CPU_WORK_BATCH queued items.
1553 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1556 int next_cpu = hctx->next_cpu;
1558 if (hctx->queue->nr_hw_queues == 1)
1559 return WORK_CPU_UNBOUND;
1561 if (--hctx->next_cpu_batch <= 0) {
1563 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1565 if (next_cpu >= nr_cpu_ids)
1566 next_cpu = blk_mq_first_mapped_cpu(hctx);
1567 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1571 * Do unbound schedule if we can't find a online CPU for this hctx,
1572 * and it should only happen in the path of handling CPU DEAD.
1574 if (!cpu_online(next_cpu)) {
1581 * Make sure to re-select CPU next time once after CPUs
1582 * in hctx->cpumask become online again.
1584 hctx->next_cpu = next_cpu;
1585 hctx->next_cpu_batch = 1;
1586 return WORK_CPU_UNBOUND;
1589 hctx->next_cpu = next_cpu;
1594 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1595 * @hctx: Pointer to the hardware queue to run.
1596 * @async: If we want to run the queue asynchronously.
1597 * @msecs: Microseconds of delay to wait before running the queue.
1599 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1600 * with a delay of @msecs.
1602 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1603 unsigned long msecs)
1605 if (unlikely(blk_mq_hctx_stopped(hctx)))
1608 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1609 int cpu = get_cpu();
1610 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1611 __blk_mq_run_hw_queue(hctx);
1619 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1620 msecs_to_jiffies(msecs));
1624 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1625 * @hctx: Pointer to the hardware queue to run.
1626 * @msecs: Microseconds of delay to wait before running the queue.
1628 * Run a hardware queue asynchronously with a delay of @msecs.
1630 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1632 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1634 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1637 * blk_mq_run_hw_queue - Start to run a hardware queue.
1638 * @hctx: Pointer to the hardware queue to run.
1639 * @async: If we want to run the queue asynchronously.
1641 * Check if the request queue is not in a quiesced state and if there are
1642 * pending requests to be sent. If this is true, run the queue to send requests
1645 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1651 * When queue is quiesced, we may be switching io scheduler, or
1652 * updating nr_hw_queues, or other things, and we can't run queue
1653 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1655 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1658 hctx_lock(hctx, &srcu_idx);
1659 need_run = !blk_queue_quiesced(hctx->queue) &&
1660 blk_mq_hctx_has_pending(hctx);
1661 hctx_unlock(hctx, srcu_idx);
1664 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1666 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1669 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1670 * @q: Pointer to the request queue to run.
1671 * @async: If we want to run the queue asynchronously.
1673 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1675 struct blk_mq_hw_ctx *hctx;
1678 queue_for_each_hw_ctx(q, hctx, i) {
1679 if (blk_mq_hctx_stopped(hctx))
1682 blk_mq_run_hw_queue(hctx, async);
1685 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1688 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1689 * @q: Pointer to the request queue to run.
1690 * @msecs: Microseconds of delay to wait before running the queues.
1692 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1694 struct blk_mq_hw_ctx *hctx;
1697 queue_for_each_hw_ctx(q, hctx, i) {
1698 if (blk_mq_hctx_stopped(hctx))
1701 blk_mq_delay_run_hw_queue(hctx, msecs);
1704 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1707 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1708 * @q: request queue.
1710 * The caller is responsible for serializing this function against
1711 * blk_mq_{start,stop}_hw_queue().
1713 bool blk_mq_queue_stopped(struct request_queue *q)
1715 struct blk_mq_hw_ctx *hctx;
1718 queue_for_each_hw_ctx(q, hctx, i)
1719 if (blk_mq_hctx_stopped(hctx))
1724 EXPORT_SYMBOL(blk_mq_queue_stopped);
1727 * This function is often used for pausing .queue_rq() by driver when
1728 * there isn't enough resource or some conditions aren't satisfied, and
1729 * BLK_STS_RESOURCE is usually returned.
1731 * We do not guarantee that dispatch can be drained or blocked
1732 * after blk_mq_stop_hw_queue() returns. Please use
1733 * blk_mq_quiesce_queue() for that requirement.
1735 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1737 cancel_delayed_work(&hctx->run_work);
1739 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1741 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1744 * This function is often used for pausing .queue_rq() by driver when
1745 * there isn't enough resource or some conditions aren't satisfied, and
1746 * BLK_STS_RESOURCE is usually returned.
1748 * We do not guarantee that dispatch can be drained or blocked
1749 * after blk_mq_stop_hw_queues() returns. Please use
1750 * blk_mq_quiesce_queue() for that requirement.
1752 void blk_mq_stop_hw_queues(struct request_queue *q)
1754 struct blk_mq_hw_ctx *hctx;
1757 queue_for_each_hw_ctx(q, hctx, i)
1758 blk_mq_stop_hw_queue(hctx);
1760 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1762 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1764 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1766 blk_mq_run_hw_queue(hctx, false);
1768 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1770 void blk_mq_start_hw_queues(struct request_queue *q)
1772 struct blk_mq_hw_ctx *hctx;
1775 queue_for_each_hw_ctx(q, hctx, i)
1776 blk_mq_start_hw_queue(hctx);
1778 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1780 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1782 if (!blk_mq_hctx_stopped(hctx))
1785 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1786 blk_mq_run_hw_queue(hctx, async);
1788 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1790 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1792 struct blk_mq_hw_ctx *hctx;
1795 queue_for_each_hw_ctx(q, hctx, i)
1796 blk_mq_start_stopped_hw_queue(hctx, async);
1798 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1800 static void blk_mq_run_work_fn(struct work_struct *work)
1802 struct blk_mq_hw_ctx *hctx;
1804 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1807 * If we are stopped, don't run the queue.
1809 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1812 __blk_mq_run_hw_queue(hctx);
1815 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1819 struct blk_mq_ctx *ctx = rq->mq_ctx;
1820 enum hctx_type type = hctx->type;
1822 lockdep_assert_held(&ctx->lock);
1824 trace_block_rq_insert(hctx->queue, rq);
1827 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1829 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1832 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1835 struct blk_mq_ctx *ctx = rq->mq_ctx;
1837 lockdep_assert_held(&ctx->lock);
1839 __blk_mq_insert_req_list(hctx, rq, at_head);
1840 blk_mq_hctx_mark_pending(hctx, ctx);
1844 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1845 * @rq: Pointer to request to be inserted.
1846 * @at_head: true if the request should be inserted at the head of the list.
1847 * @run_queue: If we should run the hardware queue after inserting the request.
1849 * Should only be used carefully, when the caller knows we want to
1850 * bypass a potential IO scheduler on the target device.
1852 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1855 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1857 spin_lock(&hctx->lock);
1859 list_add(&rq->queuelist, &hctx->dispatch);
1861 list_add_tail(&rq->queuelist, &hctx->dispatch);
1862 spin_unlock(&hctx->lock);
1865 blk_mq_run_hw_queue(hctx, false);
1868 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1869 struct list_head *list)
1873 enum hctx_type type = hctx->type;
1876 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1879 list_for_each_entry(rq, list, queuelist) {
1880 BUG_ON(rq->mq_ctx != ctx);
1881 trace_block_rq_insert(hctx->queue, rq);
1884 spin_lock(&ctx->lock);
1885 list_splice_tail_init(list, &ctx->rq_lists[type]);
1886 blk_mq_hctx_mark_pending(hctx, ctx);
1887 spin_unlock(&ctx->lock);
1890 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1892 struct request *rqa = container_of(a, struct request, queuelist);
1893 struct request *rqb = container_of(b, struct request, queuelist);
1895 if (rqa->mq_ctx != rqb->mq_ctx)
1896 return rqa->mq_ctx > rqb->mq_ctx;
1897 if (rqa->mq_hctx != rqb->mq_hctx)
1898 return rqa->mq_hctx > rqb->mq_hctx;
1900 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1903 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1907 if (list_empty(&plug->mq_list))
1909 list_splice_init(&plug->mq_list, &list);
1911 if (plug->rq_count > 2 && plug->multiple_queues)
1912 list_sort(NULL, &list, plug_rq_cmp);
1917 struct list_head rq_list;
1918 struct request *rq, *head_rq = list_entry_rq(list.next);
1919 struct list_head *pos = &head_rq->queuelist; /* skip first */
1920 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1921 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1922 unsigned int depth = 1;
1924 list_for_each_continue(pos, &list) {
1925 rq = list_entry_rq(pos);
1927 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1932 list_cut_before(&rq_list, &list, pos);
1933 trace_block_unplug(head_rq->q, depth, !from_schedule);
1934 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1936 } while(!list_empty(&list));
1939 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1940 unsigned int nr_segs)
1942 if (bio->bi_opf & REQ_RAHEAD)
1943 rq->cmd_flags |= REQ_FAILFAST_MASK;
1945 rq->__sector = bio->bi_iter.bi_sector;
1946 rq->write_hint = bio->bi_write_hint;
1947 blk_rq_bio_prep(rq, bio, nr_segs);
1948 blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1950 blk_account_io_start(rq);
1953 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1955 blk_qc_t *cookie, bool last)
1957 struct request_queue *q = rq->q;
1958 struct blk_mq_queue_data bd = {
1962 blk_qc_t new_cookie;
1965 new_cookie = request_to_qc_t(hctx, rq);
1968 * For OK queue, we are done. For error, caller may kill it.
1969 * Any other error (busy), just add it to our list as we
1970 * previously would have done.
1972 ret = q->mq_ops->queue_rq(hctx, &bd);
1975 blk_mq_update_dispatch_busy(hctx, false);
1976 *cookie = new_cookie;
1978 case BLK_STS_RESOURCE:
1979 case BLK_STS_DEV_RESOURCE:
1980 blk_mq_update_dispatch_busy(hctx, true);
1981 __blk_mq_requeue_request(rq);
1984 blk_mq_update_dispatch_busy(hctx, false);
1985 *cookie = BLK_QC_T_NONE;
1992 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1995 bool bypass_insert, bool last)
1997 struct request_queue *q = rq->q;
1998 bool run_queue = true;
2001 * RCU or SRCU read lock is needed before checking quiesced flag.
2003 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2004 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2005 * and avoid driver to try to dispatch again.
2007 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2009 bypass_insert = false;
2013 if (q->elevator && !bypass_insert)
2016 if (!blk_mq_get_dispatch_budget(q))
2019 if (!blk_mq_get_driver_tag(rq)) {
2020 blk_mq_put_dispatch_budget(q);
2024 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2027 return BLK_STS_RESOURCE;
2029 blk_mq_request_bypass_insert(rq, false, run_queue);
2034 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2035 * @hctx: Pointer of the associated hardware queue.
2036 * @rq: Pointer to request to be sent.
2037 * @cookie: Request queue cookie.
2039 * If the device has enough resources to accept a new request now, send the
2040 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2041 * we can try send it another time in the future. Requests inserted at this
2042 * queue have higher priority.
2044 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2045 struct request *rq, blk_qc_t *cookie)
2050 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2052 hctx_lock(hctx, &srcu_idx);
2054 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2055 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2056 blk_mq_request_bypass_insert(rq, false, true);
2057 else if (ret != BLK_STS_OK)
2058 blk_mq_end_request(rq, ret);
2060 hctx_unlock(hctx, srcu_idx);
2063 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2067 blk_qc_t unused_cookie;
2068 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2070 hctx_lock(hctx, &srcu_idx);
2071 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2072 hctx_unlock(hctx, srcu_idx);
2077 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2078 struct list_head *list)
2082 while (!list_empty(list)) {
2084 struct request *rq = list_first_entry(list, struct request,
2087 list_del_init(&rq->queuelist);
2088 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2089 if (ret != BLK_STS_OK) {
2090 if (ret == BLK_STS_RESOURCE ||
2091 ret == BLK_STS_DEV_RESOURCE) {
2092 blk_mq_request_bypass_insert(rq, false,
2096 blk_mq_end_request(rq, ret);
2102 * If we didn't flush the entire list, we could have told
2103 * the driver there was more coming, but that turned out to
2106 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs && queued)
2107 hctx->queue->mq_ops->commit_rqs(hctx);
2110 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2112 list_add_tail(&rq->queuelist, &plug->mq_list);
2114 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2115 struct request *tmp;
2117 tmp = list_first_entry(&plug->mq_list, struct request,
2119 if (tmp->q != rq->q)
2120 plug->multiple_queues = true;
2125 * blk_mq_submit_bio - Create and send a request to block device.
2126 * @bio: Bio pointer.
2128 * Builds up a request structure from @q and @bio and send to the device. The
2129 * request may not be queued directly to hardware if:
2130 * * This request can be merged with another one
2131 * * We want to place request at plug queue for possible future merging
2132 * * There is an IO scheduler active at this queue
2134 * It will not queue the request if there is an error with the bio, or at the
2137 * Returns: Request queue cookie.
2139 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2141 struct request_queue *q = bio->bi_disk->queue;
2142 const int is_sync = op_is_sync(bio->bi_opf);
2143 const int is_flush_fua = op_is_flush(bio->bi_opf);
2144 struct blk_mq_alloc_data data = {
2148 struct blk_plug *plug;
2149 struct request *same_queue_rq = NULL;
2150 unsigned int nr_segs;
2154 blk_queue_bounce(q, &bio);
2155 __blk_queue_split(&bio, &nr_segs);
2157 if (!bio_integrity_prep(bio))
2160 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2161 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2164 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2167 rq_qos_throttle(q, bio);
2169 data.cmd_flags = bio->bi_opf;
2170 rq = __blk_mq_alloc_request(&data);
2171 if (unlikely(!rq)) {
2172 rq_qos_cleanup(q, bio);
2173 if (bio->bi_opf & REQ_NOWAIT)
2174 bio_wouldblock_error(bio);
2178 trace_block_getrq(q, bio, bio->bi_opf);
2180 rq_qos_track(q, rq, bio);
2182 cookie = request_to_qc_t(data.hctx, rq);
2184 blk_mq_bio_to_request(rq, bio, nr_segs);
2186 ret = blk_crypto_init_request(rq);
2187 if (ret != BLK_STS_OK) {
2188 bio->bi_status = ret;
2190 blk_mq_free_request(rq);
2191 return BLK_QC_T_NONE;
2194 plug = blk_mq_plug(q, bio);
2195 if (unlikely(is_flush_fua)) {
2196 /* Bypass scheduler for flush requests */
2197 blk_insert_flush(rq);
2198 blk_mq_run_hw_queue(data.hctx, true);
2199 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2200 !blk_queue_nonrot(q))) {
2202 * Use plugging if we have a ->commit_rqs() hook as well, as
2203 * we know the driver uses bd->last in a smart fashion.
2205 * Use normal plugging if this disk is slow HDD, as sequential
2206 * IO may benefit a lot from plug merging.
2208 unsigned int request_count = plug->rq_count;
2209 struct request *last = NULL;
2212 trace_block_plug(q);
2214 last = list_entry_rq(plug->mq_list.prev);
2216 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2217 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2218 blk_flush_plug_list(plug, false);
2219 trace_block_plug(q);
2222 blk_add_rq_to_plug(plug, rq);
2223 } else if (q->elevator) {
2224 /* Insert the request at the IO scheduler queue */
2225 blk_mq_sched_insert_request(rq, false, true, true);
2226 } else if (plug && !blk_queue_nomerges(q)) {
2228 * We do limited plugging. If the bio can be merged, do that.
2229 * Otherwise the existing request in the plug list will be
2230 * issued. So the plug list will have one request at most
2231 * The plug list might get flushed before this. If that happens,
2232 * the plug list is empty, and same_queue_rq is invalid.
2234 if (list_empty(&plug->mq_list))
2235 same_queue_rq = NULL;
2236 if (same_queue_rq) {
2237 list_del_init(&same_queue_rq->queuelist);
2240 blk_add_rq_to_plug(plug, rq);
2241 trace_block_plug(q);
2243 if (same_queue_rq) {
2244 data.hctx = same_queue_rq->mq_hctx;
2245 trace_block_unplug(q, 1, true);
2246 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2249 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2250 !data.hctx->dispatch_busy) {
2252 * There is no scheduler and we can try to send directly
2255 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2258 blk_mq_sched_insert_request(rq, false, true, true);
2264 return BLK_QC_T_NONE;
2266 EXPORT_SYMBOL_GPL(blk_mq_submit_bio); /* only for request based dm */
2268 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2269 unsigned int hctx_idx)
2273 if (tags->rqs && set->ops->exit_request) {
2276 for (i = 0; i < tags->nr_tags; i++) {
2277 struct request *rq = tags->static_rqs[i];
2281 set->ops->exit_request(set, rq, hctx_idx);
2282 tags->static_rqs[i] = NULL;
2286 while (!list_empty(&tags->page_list)) {
2287 page = list_first_entry(&tags->page_list, struct page, lru);
2288 list_del_init(&page->lru);
2290 * Remove kmemleak object previously allocated in
2291 * blk_mq_alloc_rqs().
2293 kmemleak_free(page_address(page));
2294 __free_pages(page, page->private);
2298 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2302 kfree(tags->static_rqs);
2303 tags->static_rqs = NULL;
2305 blk_mq_free_tags(tags);
2308 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2309 unsigned int hctx_idx,
2310 unsigned int nr_tags,
2311 unsigned int reserved_tags)
2313 struct blk_mq_tags *tags;
2316 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2317 if (node == NUMA_NO_NODE)
2318 node = set->numa_node;
2320 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2321 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2325 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2326 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2329 blk_mq_free_tags(tags);
2333 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2334 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2336 if (!tags->static_rqs) {
2338 blk_mq_free_tags(tags);
2345 static size_t order_to_size(unsigned int order)
2347 return (size_t)PAGE_SIZE << order;
2350 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2351 unsigned int hctx_idx, int node)
2355 if (set->ops->init_request) {
2356 ret = set->ops->init_request(set, rq, hctx_idx, node);
2361 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2365 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2366 unsigned int hctx_idx, unsigned int depth)
2368 unsigned int i, j, entries_per_page, max_order = 4;
2369 size_t rq_size, left;
2372 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2373 if (node == NUMA_NO_NODE)
2374 node = set->numa_node;
2376 INIT_LIST_HEAD(&tags->page_list);
2379 * rq_size is the size of the request plus driver payload, rounded
2380 * to the cacheline size
2382 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2384 left = rq_size * depth;
2386 for (i = 0; i < depth; ) {
2387 int this_order = max_order;
2392 while (this_order && left < order_to_size(this_order - 1))
2396 page = alloc_pages_node(node,
2397 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2403 if (order_to_size(this_order) < rq_size)
2410 page->private = this_order;
2411 list_add_tail(&page->lru, &tags->page_list);
2413 p = page_address(page);
2415 * Allow kmemleak to scan these pages as they contain pointers
2416 * to additional allocations like via ops->init_request().
2418 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2419 entries_per_page = order_to_size(this_order) / rq_size;
2420 to_do = min(entries_per_page, depth - i);
2421 left -= to_do * rq_size;
2422 for (j = 0; j < to_do; j++) {
2423 struct request *rq = p;
2425 tags->static_rqs[i] = rq;
2426 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2427 tags->static_rqs[i] = NULL;
2438 blk_mq_free_rqs(set, tags, hctx_idx);
2442 struct rq_iter_data {
2443 struct blk_mq_hw_ctx *hctx;
2447 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2449 struct rq_iter_data *iter_data = data;
2451 if (rq->mq_hctx != iter_data->hctx)
2453 iter_data->has_rq = true;
2457 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2459 struct blk_mq_tags *tags = hctx->sched_tags ?
2460 hctx->sched_tags : hctx->tags;
2461 struct rq_iter_data data = {
2465 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2469 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2470 struct blk_mq_hw_ctx *hctx)
2472 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2474 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2479 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2481 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2482 struct blk_mq_hw_ctx, cpuhp_online);
2484 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2485 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2489 * Prevent new request from being allocated on the current hctx.
2491 * The smp_mb__after_atomic() Pairs with the implied barrier in
2492 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2493 * seen once we return from the tag allocator.
2495 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2496 smp_mb__after_atomic();
2499 * Try to grab a reference to the queue and wait for any outstanding
2500 * requests. If we could not grab a reference the queue has been
2501 * frozen and there are no requests.
2503 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2504 while (blk_mq_hctx_has_requests(hctx))
2506 percpu_ref_put(&hctx->queue->q_usage_counter);
2512 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2514 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2515 struct blk_mq_hw_ctx, cpuhp_online);
2517 if (cpumask_test_cpu(cpu, hctx->cpumask))
2518 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2523 * 'cpu' is going away. splice any existing rq_list entries from this
2524 * software queue to the hw queue dispatch list, and ensure that it
2527 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2529 struct blk_mq_hw_ctx *hctx;
2530 struct blk_mq_ctx *ctx;
2532 enum hctx_type type;
2534 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2535 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2538 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2541 spin_lock(&ctx->lock);
2542 if (!list_empty(&ctx->rq_lists[type])) {
2543 list_splice_init(&ctx->rq_lists[type], &tmp);
2544 blk_mq_hctx_clear_pending(hctx, ctx);
2546 spin_unlock(&ctx->lock);
2548 if (list_empty(&tmp))
2551 spin_lock(&hctx->lock);
2552 list_splice_tail_init(&tmp, &hctx->dispatch);
2553 spin_unlock(&hctx->lock);
2555 blk_mq_run_hw_queue(hctx, true);
2559 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2561 if (!(hctx->flags & BLK_MQ_F_STACKING))
2562 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2563 &hctx->cpuhp_online);
2564 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2568 /* hctx->ctxs will be freed in queue's release handler */
2569 static void blk_mq_exit_hctx(struct request_queue *q,
2570 struct blk_mq_tag_set *set,
2571 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2573 if (blk_mq_hw_queue_mapped(hctx))
2574 blk_mq_tag_idle(hctx);
2576 if (set->ops->exit_request)
2577 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2579 if (set->ops->exit_hctx)
2580 set->ops->exit_hctx(hctx, hctx_idx);
2582 blk_mq_remove_cpuhp(hctx);
2584 spin_lock(&q->unused_hctx_lock);
2585 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2586 spin_unlock(&q->unused_hctx_lock);
2589 static void blk_mq_exit_hw_queues(struct request_queue *q,
2590 struct blk_mq_tag_set *set, int nr_queue)
2592 struct blk_mq_hw_ctx *hctx;
2595 queue_for_each_hw_ctx(q, hctx, i) {
2598 blk_mq_debugfs_unregister_hctx(hctx);
2599 blk_mq_exit_hctx(q, set, hctx, i);
2603 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2605 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2607 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2608 __alignof__(struct blk_mq_hw_ctx)) !=
2609 sizeof(struct blk_mq_hw_ctx));
2611 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2612 hw_ctx_size += sizeof(struct srcu_struct);
2617 static int blk_mq_init_hctx(struct request_queue *q,
2618 struct blk_mq_tag_set *set,
2619 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2621 hctx->queue_num = hctx_idx;
2623 if (!(hctx->flags & BLK_MQ_F_STACKING))
2624 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2625 &hctx->cpuhp_online);
2626 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2628 hctx->tags = set->tags[hctx_idx];
2630 if (set->ops->init_hctx &&
2631 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2632 goto unregister_cpu_notifier;
2634 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2640 if (set->ops->exit_hctx)
2641 set->ops->exit_hctx(hctx, hctx_idx);
2642 unregister_cpu_notifier:
2643 blk_mq_remove_cpuhp(hctx);
2647 static struct blk_mq_hw_ctx *
2648 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2651 struct blk_mq_hw_ctx *hctx;
2652 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2654 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2656 goto fail_alloc_hctx;
2658 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2661 atomic_set(&hctx->nr_active, 0);
2662 if (node == NUMA_NO_NODE)
2663 node = set->numa_node;
2664 hctx->numa_node = node;
2666 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2667 spin_lock_init(&hctx->lock);
2668 INIT_LIST_HEAD(&hctx->dispatch);
2670 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2672 INIT_LIST_HEAD(&hctx->hctx_list);
2675 * Allocate space for all possible cpus to avoid allocation at
2678 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2683 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2688 spin_lock_init(&hctx->dispatch_wait_lock);
2689 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2690 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2692 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2696 if (hctx->flags & BLK_MQ_F_BLOCKING)
2697 init_srcu_struct(hctx->srcu);
2698 blk_mq_hctx_kobj_init(hctx);
2703 sbitmap_free(&hctx->ctx_map);
2707 free_cpumask_var(hctx->cpumask);
2714 static void blk_mq_init_cpu_queues(struct request_queue *q,
2715 unsigned int nr_hw_queues)
2717 struct blk_mq_tag_set *set = q->tag_set;
2720 for_each_possible_cpu(i) {
2721 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2722 struct blk_mq_hw_ctx *hctx;
2726 spin_lock_init(&__ctx->lock);
2727 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2728 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2733 * Set local node, IFF we have more than one hw queue. If
2734 * not, we remain on the home node of the device
2736 for (j = 0; j < set->nr_maps; j++) {
2737 hctx = blk_mq_map_queue_type(q, j, i);
2738 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2739 hctx->numa_node = local_memory_node(cpu_to_node(i));
2744 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2749 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2750 set->queue_depth, set->reserved_tags);
2751 if (!set->tags[hctx_idx])
2754 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2759 blk_mq_free_rq_map(set->tags[hctx_idx]);
2760 set->tags[hctx_idx] = NULL;
2764 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2765 unsigned int hctx_idx)
2767 if (set->tags && set->tags[hctx_idx]) {
2768 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2769 blk_mq_free_rq_map(set->tags[hctx_idx]);
2770 set->tags[hctx_idx] = NULL;
2774 static void blk_mq_map_swqueue(struct request_queue *q)
2776 unsigned int i, j, hctx_idx;
2777 struct blk_mq_hw_ctx *hctx;
2778 struct blk_mq_ctx *ctx;
2779 struct blk_mq_tag_set *set = q->tag_set;
2781 queue_for_each_hw_ctx(q, hctx, i) {
2782 cpumask_clear(hctx->cpumask);
2784 hctx->dispatch_from = NULL;
2788 * Map software to hardware queues.
2790 * If the cpu isn't present, the cpu is mapped to first hctx.
2792 for_each_possible_cpu(i) {
2794 ctx = per_cpu_ptr(q->queue_ctx, i);
2795 for (j = 0; j < set->nr_maps; j++) {
2796 if (!set->map[j].nr_queues) {
2797 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2798 HCTX_TYPE_DEFAULT, i);
2801 hctx_idx = set->map[j].mq_map[i];
2802 /* unmapped hw queue can be remapped after CPU topo changed */
2803 if (!set->tags[hctx_idx] &&
2804 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2806 * If tags initialization fail for some hctx,
2807 * that hctx won't be brought online. In this
2808 * case, remap the current ctx to hctx[0] which
2809 * is guaranteed to always have tags allocated
2811 set->map[j].mq_map[i] = 0;
2814 hctx = blk_mq_map_queue_type(q, j, i);
2815 ctx->hctxs[j] = hctx;
2817 * If the CPU is already set in the mask, then we've
2818 * mapped this one already. This can happen if
2819 * devices share queues across queue maps.
2821 if (cpumask_test_cpu(i, hctx->cpumask))
2824 cpumask_set_cpu(i, hctx->cpumask);
2826 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2827 hctx->ctxs[hctx->nr_ctx++] = ctx;
2830 * If the nr_ctx type overflows, we have exceeded the
2831 * amount of sw queues we can support.
2833 BUG_ON(!hctx->nr_ctx);
2836 for (; j < HCTX_MAX_TYPES; j++)
2837 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2838 HCTX_TYPE_DEFAULT, i);
2841 queue_for_each_hw_ctx(q, hctx, i) {
2843 * If no software queues are mapped to this hardware queue,
2844 * disable it and free the request entries.
2846 if (!hctx->nr_ctx) {
2847 /* Never unmap queue 0. We need it as a
2848 * fallback in case of a new remap fails
2851 if (i && set->tags[i])
2852 blk_mq_free_map_and_requests(set, i);
2858 hctx->tags = set->tags[i];
2859 WARN_ON(!hctx->tags);
2862 * Set the map size to the number of mapped software queues.
2863 * This is more accurate and more efficient than looping
2864 * over all possibly mapped software queues.
2866 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2869 * Initialize batch roundrobin counts
2871 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2872 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2877 * Caller needs to ensure that we're either frozen/quiesced, or that
2878 * the queue isn't live yet.
2880 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2882 struct blk_mq_hw_ctx *hctx;
2885 queue_for_each_hw_ctx(q, hctx, i) {
2887 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2889 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2893 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2896 struct request_queue *q;
2898 lockdep_assert_held(&set->tag_list_lock);
2900 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2901 blk_mq_freeze_queue(q);
2902 queue_set_hctx_shared(q, shared);
2903 blk_mq_unfreeze_queue(q);
2907 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2909 struct blk_mq_tag_set *set = q->tag_set;
2911 mutex_lock(&set->tag_list_lock);
2912 list_del(&q->tag_set_list);
2913 if (list_is_singular(&set->tag_list)) {
2914 /* just transitioned to unshared */
2915 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2916 /* update existing queue */
2917 blk_mq_update_tag_set_depth(set, false);
2919 mutex_unlock(&set->tag_list_lock);
2920 INIT_LIST_HEAD(&q->tag_set_list);
2923 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2924 struct request_queue *q)
2926 mutex_lock(&set->tag_list_lock);
2929 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2931 if (!list_empty(&set->tag_list) &&
2932 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2933 set->flags |= BLK_MQ_F_TAG_SHARED;
2934 /* update existing queue */
2935 blk_mq_update_tag_set_depth(set, true);
2937 if (set->flags & BLK_MQ_F_TAG_SHARED)
2938 queue_set_hctx_shared(q, true);
2939 list_add_tail(&q->tag_set_list, &set->tag_list);
2941 mutex_unlock(&set->tag_list_lock);
2944 /* All allocations will be freed in release handler of q->mq_kobj */
2945 static int blk_mq_alloc_ctxs(struct request_queue *q)
2947 struct blk_mq_ctxs *ctxs;
2950 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2954 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2955 if (!ctxs->queue_ctx)
2958 for_each_possible_cpu(cpu) {
2959 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2963 q->mq_kobj = &ctxs->kobj;
2964 q->queue_ctx = ctxs->queue_ctx;
2973 * It is the actual release handler for mq, but we do it from
2974 * request queue's release handler for avoiding use-after-free
2975 * and headache because q->mq_kobj shouldn't have been introduced,
2976 * but we can't group ctx/kctx kobj without it.
2978 void blk_mq_release(struct request_queue *q)
2980 struct blk_mq_hw_ctx *hctx, *next;
2983 queue_for_each_hw_ctx(q, hctx, i)
2984 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2986 /* all hctx are in .unused_hctx_list now */
2987 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2988 list_del_init(&hctx->hctx_list);
2989 kobject_put(&hctx->kobj);
2992 kfree(q->queue_hw_ctx);
2995 * release .mq_kobj and sw queue's kobject now because
2996 * both share lifetime with request queue.
2998 blk_mq_sysfs_deinit(q);
3001 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3004 struct request_queue *uninit_q, *q;
3006 uninit_q = blk_alloc_queue(set->numa_node);
3008 return ERR_PTR(-ENOMEM);
3009 uninit_q->queuedata = queuedata;
3012 * Initialize the queue without an elevator. device_add_disk() will do
3013 * the initialization.
3015 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3017 blk_cleanup_queue(uninit_q);
3021 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3023 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3025 return blk_mq_init_queue_data(set, NULL);
3027 EXPORT_SYMBOL(blk_mq_init_queue);
3030 * Helper for setting up a queue with mq ops, given queue depth, and
3031 * the passed in mq ops flags.
3033 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3034 const struct blk_mq_ops *ops,
3035 unsigned int queue_depth,
3036 unsigned int set_flags)
3038 struct request_queue *q;
3041 memset(set, 0, sizeof(*set));
3043 set->nr_hw_queues = 1;
3045 set->queue_depth = queue_depth;
3046 set->numa_node = NUMA_NO_NODE;
3047 set->flags = set_flags;
3049 ret = blk_mq_alloc_tag_set(set);
3051 return ERR_PTR(ret);
3053 q = blk_mq_init_queue(set);
3055 blk_mq_free_tag_set(set);
3061 EXPORT_SYMBOL(blk_mq_init_sq_queue);
3063 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3064 struct blk_mq_tag_set *set, struct request_queue *q,
3065 int hctx_idx, int node)
3067 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3069 /* reuse dead hctx first */
3070 spin_lock(&q->unused_hctx_lock);
3071 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3072 if (tmp->numa_node == node) {
3078 list_del_init(&hctx->hctx_list);
3079 spin_unlock(&q->unused_hctx_lock);
3082 hctx = blk_mq_alloc_hctx(q, set, node);
3086 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3092 kobject_put(&hctx->kobj);
3097 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3098 struct request_queue *q)
3101 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3103 if (q->nr_hw_queues < set->nr_hw_queues) {
3104 struct blk_mq_hw_ctx **new_hctxs;
3106 new_hctxs = kcalloc_node(set->nr_hw_queues,
3107 sizeof(*new_hctxs), GFP_KERNEL,
3112 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3114 q->queue_hw_ctx = new_hctxs;
3119 /* protect against switching io scheduler */
3120 mutex_lock(&q->sysfs_lock);
3121 for (i = 0; i < set->nr_hw_queues; i++) {
3123 struct blk_mq_hw_ctx *hctx;
3125 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3127 * If the hw queue has been mapped to another numa node,
3128 * we need to realloc the hctx. If allocation fails, fallback
3129 * to use the previous one.
3131 if (hctxs[i] && (hctxs[i]->numa_node == node))
3134 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3137 blk_mq_exit_hctx(q, set, hctxs[i], i);
3141 pr_warn("Allocate new hctx on node %d fails,\
3142 fallback to previous one on node %d\n",
3143 node, hctxs[i]->numa_node);
3149 * Increasing nr_hw_queues fails. Free the newly allocated
3150 * hctxs and keep the previous q->nr_hw_queues.
3152 if (i != set->nr_hw_queues) {
3153 j = q->nr_hw_queues;
3157 end = q->nr_hw_queues;
3158 q->nr_hw_queues = set->nr_hw_queues;
3161 for (; j < end; j++) {
3162 struct blk_mq_hw_ctx *hctx = hctxs[j];
3166 blk_mq_free_map_and_requests(set, j);
3167 blk_mq_exit_hctx(q, set, hctx, j);
3171 mutex_unlock(&q->sysfs_lock);
3174 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3175 struct request_queue *q,
3178 /* mark the queue as mq asap */
3179 q->mq_ops = set->ops;
3181 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3182 blk_mq_poll_stats_bkt,
3183 BLK_MQ_POLL_STATS_BKTS, q);
3187 if (blk_mq_alloc_ctxs(q))
3190 /* init q->mq_kobj and sw queues' kobjects */
3191 blk_mq_sysfs_init(q);
3193 INIT_LIST_HEAD(&q->unused_hctx_list);
3194 spin_lock_init(&q->unused_hctx_lock);
3196 blk_mq_realloc_hw_ctxs(set, q);
3197 if (!q->nr_hw_queues)
3200 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3201 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3205 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3206 if (set->nr_maps > HCTX_TYPE_POLL &&
3207 set->map[HCTX_TYPE_POLL].nr_queues)
3208 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3210 q->sg_reserved_size = INT_MAX;
3212 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3213 INIT_LIST_HEAD(&q->requeue_list);
3214 spin_lock_init(&q->requeue_lock);
3216 q->nr_requests = set->queue_depth;
3219 * Default to classic polling
3221 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3223 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3224 blk_mq_add_queue_tag_set(set, q);
3225 blk_mq_map_swqueue(q);
3228 elevator_init_mq(q);
3233 kfree(q->queue_hw_ctx);
3234 q->nr_hw_queues = 0;
3235 blk_mq_sysfs_deinit(q);
3237 blk_stat_free_callback(q->poll_cb);
3241 return ERR_PTR(-ENOMEM);
3243 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3245 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3246 void blk_mq_exit_queue(struct request_queue *q)
3248 struct blk_mq_tag_set *set = q->tag_set;
3250 blk_mq_del_queue_tag_set(q);
3251 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3254 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3258 for (i = 0; i < set->nr_hw_queues; i++)
3259 if (!__blk_mq_alloc_map_and_request(set, i))
3266 blk_mq_free_map_and_requests(set, i);
3272 * Allocate the request maps associated with this tag_set. Note that this
3273 * may reduce the depth asked for, if memory is tight. set->queue_depth
3274 * will be updated to reflect the allocated depth.
3276 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3281 depth = set->queue_depth;
3283 err = __blk_mq_alloc_rq_maps(set);
3287 set->queue_depth >>= 1;
3288 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3292 } while (set->queue_depth);
3294 if (!set->queue_depth || err) {
3295 pr_err("blk-mq: failed to allocate request map\n");
3299 if (depth != set->queue_depth)
3300 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3301 depth, set->queue_depth);
3306 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3309 * blk_mq_map_queues() and multiple .map_queues() implementations
3310 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3311 * number of hardware queues.
3313 if (set->nr_maps == 1)
3314 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3316 if (set->ops->map_queues && !is_kdump_kernel()) {
3320 * transport .map_queues is usually done in the following
3323 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3324 * mask = get_cpu_mask(queue)
3325 * for_each_cpu(cpu, mask)
3326 * set->map[x].mq_map[cpu] = queue;
3329 * When we need to remap, the table has to be cleared for
3330 * killing stale mapping since one CPU may not be mapped
3333 for (i = 0; i < set->nr_maps; i++)
3334 blk_mq_clear_mq_map(&set->map[i]);
3336 return set->ops->map_queues(set);
3338 BUG_ON(set->nr_maps > 1);
3339 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3343 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3344 int cur_nr_hw_queues, int new_nr_hw_queues)
3346 struct blk_mq_tags **new_tags;
3348 if (cur_nr_hw_queues >= new_nr_hw_queues)
3351 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3352 GFP_KERNEL, set->numa_node);
3357 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3358 sizeof(*set->tags));
3360 set->tags = new_tags;
3361 set->nr_hw_queues = new_nr_hw_queues;
3367 * Alloc a tag set to be associated with one or more request queues.
3368 * May fail with EINVAL for various error conditions. May adjust the
3369 * requested depth down, if it's too large. In that case, the set
3370 * value will be stored in set->queue_depth.
3372 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3376 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3378 if (!set->nr_hw_queues)
3380 if (!set->queue_depth)
3382 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3385 if (!set->ops->queue_rq)
3388 if (!set->ops->get_budget ^ !set->ops->put_budget)
3391 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3392 pr_info("blk-mq: reduced tag depth to %u\n",
3394 set->queue_depth = BLK_MQ_MAX_DEPTH;
3399 else if (set->nr_maps > HCTX_MAX_TYPES)
3403 * If a crashdump is active, then we are potentially in a very
3404 * memory constrained environment. Limit us to 1 queue and
3405 * 64 tags to prevent using too much memory.
3407 if (is_kdump_kernel()) {
3408 set->nr_hw_queues = 1;
3410 set->queue_depth = min(64U, set->queue_depth);
3413 * There is no use for more h/w queues than cpus if we just have
3416 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3417 set->nr_hw_queues = nr_cpu_ids;
3419 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3423 for (i = 0; i < set->nr_maps; i++) {
3424 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3425 sizeof(set->map[i].mq_map[0]),
3426 GFP_KERNEL, set->numa_node);
3427 if (!set->map[i].mq_map)
3428 goto out_free_mq_map;
3429 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3432 ret = blk_mq_update_queue_map(set);
3434 goto out_free_mq_map;
3436 ret = blk_mq_alloc_map_and_requests(set);
3438 goto out_free_mq_map;
3440 mutex_init(&set->tag_list_lock);
3441 INIT_LIST_HEAD(&set->tag_list);
3446 for (i = 0; i < set->nr_maps; i++) {
3447 kfree(set->map[i].mq_map);
3448 set->map[i].mq_map = NULL;
3454 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3456 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3460 for (i = 0; i < set->nr_hw_queues; i++)
3461 blk_mq_free_map_and_requests(set, i);
3463 for (j = 0; j < set->nr_maps; j++) {
3464 kfree(set->map[j].mq_map);
3465 set->map[j].mq_map = NULL;
3471 EXPORT_SYMBOL(blk_mq_free_tag_set);
3473 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3475 struct blk_mq_tag_set *set = q->tag_set;
3476 struct blk_mq_hw_ctx *hctx;
3482 if (q->nr_requests == nr)
3485 blk_mq_freeze_queue(q);
3486 blk_mq_quiesce_queue(q);
3489 queue_for_each_hw_ctx(q, hctx, i) {
3493 * If we're using an MQ scheduler, just update the scheduler
3494 * queue depth. This is similar to what the old code would do.
3496 if (!hctx->sched_tags) {
3497 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3500 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3505 if (q->elevator && q->elevator->type->ops.depth_updated)
3506 q->elevator->type->ops.depth_updated(hctx);
3510 q->nr_requests = nr;
3512 blk_mq_unquiesce_queue(q);
3513 blk_mq_unfreeze_queue(q);
3519 * request_queue and elevator_type pair.
3520 * It is just used by __blk_mq_update_nr_hw_queues to cache
3521 * the elevator_type associated with a request_queue.
3523 struct blk_mq_qe_pair {
3524 struct list_head node;
3525 struct request_queue *q;
3526 struct elevator_type *type;
3530 * Cache the elevator_type in qe pair list and switch the
3531 * io scheduler to 'none'
3533 static bool blk_mq_elv_switch_none(struct list_head *head,
3534 struct request_queue *q)
3536 struct blk_mq_qe_pair *qe;
3541 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3545 INIT_LIST_HEAD(&qe->node);
3547 qe->type = q->elevator->type;
3548 list_add(&qe->node, head);
3550 mutex_lock(&q->sysfs_lock);
3552 * After elevator_switch_mq, the previous elevator_queue will be
3553 * released by elevator_release. The reference of the io scheduler
3554 * module get by elevator_get will also be put. So we need to get
3555 * a reference of the io scheduler module here to prevent it to be
3558 __module_get(qe->type->elevator_owner);
3559 elevator_switch_mq(q, NULL);
3560 mutex_unlock(&q->sysfs_lock);
3565 static void blk_mq_elv_switch_back(struct list_head *head,
3566 struct request_queue *q)
3568 struct blk_mq_qe_pair *qe;
3569 struct elevator_type *t = NULL;
3571 list_for_each_entry(qe, head, node)
3580 list_del(&qe->node);
3583 mutex_lock(&q->sysfs_lock);
3584 elevator_switch_mq(q, t);
3585 mutex_unlock(&q->sysfs_lock);
3588 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3591 struct request_queue *q;
3593 int prev_nr_hw_queues;
3595 lockdep_assert_held(&set->tag_list_lock);
3597 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3598 nr_hw_queues = nr_cpu_ids;
3599 if (nr_hw_queues < 1)
3601 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3604 list_for_each_entry(q, &set->tag_list, tag_set_list)
3605 blk_mq_freeze_queue(q);
3607 * Switch IO scheduler to 'none', cleaning up the data associated
3608 * with the previous scheduler. We will switch back once we are done
3609 * updating the new sw to hw queue mappings.
3611 list_for_each_entry(q, &set->tag_list, tag_set_list)
3612 if (!blk_mq_elv_switch_none(&head, q))
3615 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3616 blk_mq_debugfs_unregister_hctxs(q);
3617 blk_mq_sysfs_unregister(q);
3620 prev_nr_hw_queues = set->nr_hw_queues;
3621 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3625 set->nr_hw_queues = nr_hw_queues;
3627 blk_mq_update_queue_map(set);
3628 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3629 blk_mq_realloc_hw_ctxs(set, q);
3630 if (q->nr_hw_queues != set->nr_hw_queues) {
3631 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3632 nr_hw_queues, prev_nr_hw_queues);
3633 set->nr_hw_queues = prev_nr_hw_queues;
3634 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3637 blk_mq_map_swqueue(q);
3641 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3642 blk_mq_sysfs_register(q);
3643 blk_mq_debugfs_register_hctxs(q);
3647 list_for_each_entry(q, &set->tag_list, tag_set_list)
3648 blk_mq_elv_switch_back(&head, q);
3650 list_for_each_entry(q, &set->tag_list, tag_set_list)
3651 blk_mq_unfreeze_queue(q);
3654 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3656 mutex_lock(&set->tag_list_lock);
3657 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3658 mutex_unlock(&set->tag_list_lock);
3660 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3662 /* Enable polling stats and return whether they were already enabled. */
3663 static bool blk_poll_stats_enable(struct request_queue *q)
3665 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3666 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3668 blk_stat_add_callback(q, q->poll_cb);
3672 static void blk_mq_poll_stats_start(struct request_queue *q)
3675 * We don't arm the callback if polling stats are not enabled or the
3676 * callback is already active.
3678 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3679 blk_stat_is_active(q->poll_cb))
3682 blk_stat_activate_msecs(q->poll_cb, 100);
3685 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3687 struct request_queue *q = cb->data;
3690 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3691 if (cb->stat[bucket].nr_samples)
3692 q->poll_stat[bucket] = cb->stat[bucket];
3696 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3699 unsigned long ret = 0;
3703 * If stats collection isn't on, don't sleep but turn it on for
3706 if (!blk_poll_stats_enable(q))
3710 * As an optimistic guess, use half of the mean service time
3711 * for this type of request. We can (and should) make this smarter.
3712 * For instance, if the completion latencies are tight, we can
3713 * get closer than just half the mean. This is especially
3714 * important on devices where the completion latencies are longer
3715 * than ~10 usec. We do use the stats for the relevant IO size
3716 * if available which does lead to better estimates.
3718 bucket = blk_mq_poll_stats_bkt(rq);
3722 if (q->poll_stat[bucket].nr_samples)
3723 ret = (q->poll_stat[bucket].mean + 1) / 2;
3728 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3731 struct hrtimer_sleeper hs;
3732 enum hrtimer_mode mode;
3736 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3740 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3742 * 0: use half of prev avg
3743 * >0: use this specific value
3745 if (q->poll_nsec > 0)
3746 nsecs = q->poll_nsec;
3748 nsecs = blk_mq_poll_nsecs(q, rq);
3753 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3756 * This will be replaced with the stats tracking code, using
3757 * 'avg_completion_time / 2' as the pre-sleep target.
3761 mode = HRTIMER_MODE_REL;
3762 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3763 hrtimer_set_expires(&hs.timer, kt);
3766 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3768 set_current_state(TASK_UNINTERRUPTIBLE);
3769 hrtimer_sleeper_start_expires(&hs, mode);
3772 hrtimer_cancel(&hs.timer);
3773 mode = HRTIMER_MODE_ABS;
3774 } while (hs.task && !signal_pending(current));
3776 __set_current_state(TASK_RUNNING);
3777 destroy_hrtimer_on_stack(&hs.timer);
3781 static bool blk_mq_poll_hybrid(struct request_queue *q,
3782 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3786 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3789 if (!blk_qc_t_is_internal(cookie))
3790 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3792 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3794 * With scheduling, if the request has completed, we'll
3795 * get a NULL return here, as we clear the sched tag when
3796 * that happens. The request still remains valid, like always,
3797 * so we should be safe with just the NULL check.
3803 return blk_mq_poll_hybrid_sleep(q, rq);
3807 * blk_poll - poll for IO completions
3809 * @cookie: cookie passed back at IO submission time
3810 * @spin: whether to spin for completions
3813 * Poll for completions on the passed in queue. Returns number of
3814 * completed entries found. If @spin is true, then blk_poll will continue
3815 * looping until at least one completion is found, unless the task is
3816 * otherwise marked running (or we need to reschedule).
3818 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3820 struct blk_mq_hw_ctx *hctx;
3823 if (!blk_qc_t_valid(cookie) ||
3824 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3828 blk_flush_plug_list(current->plug, false);
3830 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3833 * If we sleep, have the caller restart the poll loop to reset
3834 * the state. Like for the other success return cases, the
3835 * caller is responsible for checking if the IO completed. If
3836 * the IO isn't complete, we'll get called again and will go
3837 * straight to the busy poll loop.
3839 if (blk_mq_poll_hybrid(q, hctx, cookie))
3842 hctx->poll_considered++;
3844 state = current->state;
3848 hctx->poll_invoked++;
3850 ret = q->mq_ops->poll(hctx);
3852 hctx->poll_success++;
3853 __set_current_state(TASK_RUNNING);
3857 if (signal_pending_state(state, current))
3858 __set_current_state(TASK_RUNNING);
3860 if (current->state == TASK_RUNNING)
3862 if (ret < 0 || !spin)
3865 } while (!need_resched());
3867 __set_current_state(TASK_RUNNING);
3870 EXPORT_SYMBOL_GPL(blk_poll);
3872 unsigned int blk_mq_rq_cpu(struct request *rq)
3874 return rq->mq_ctx->cpu;
3876 EXPORT_SYMBOL(blk_mq_rq_cpu);
3878 static int __init blk_mq_init(void)
3882 for_each_possible_cpu(i)
3883 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3884 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3886 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3887 "block/softirq:dead", NULL,
3888 blk_softirq_cpu_dead);
3889 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3890 blk_mq_hctx_notify_dead);
3891 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3892 blk_mq_hctx_notify_online,
3893 blk_mq_hctx_notify_offline);
3896 subsys_initcall(blk_mq_init);